Documento completo qui
Marco Castelletti
Turbulence models for the Reynolds-averaged Navier-Stokes equations (RANS) based on the eddy viscosity concept still are the leading approach for Computational Fluid Dynamics (CFD) simulations. Nevertheless, universal models with good predictive capabilities over a wide range of flows remain a challenge.In this work, machine learning is used to enhance the ability of existing turbulence models to provide acceptable results. The approach leverages a neural network to identify and segment different zones in the flow field, and locally adapts the modeling of turbulence to the physical nature of each zone, via a proper combination of existing models that have been preliminarily tuned to work well in a small set of elementary flows.
The work presents the idea, accompanied by a preliminary implementation of the methodology, where only two different flow zones are identified, and only two standard turbulence models are used. A test case demonstrates that, already in this oversimplified form, predictive capabilities are improved in comparison to the baseline RANS models.